Context Navigation

heap.cfa @ 3e4bf0d

Last change on this file since 3e4bf0d was 1e7a765, checked in by Peter A. Buhr <pabuhr@…>, 9 months ago
fix lock/unlock for heap statistic-mode
Property mode set to `100644`
File size: 68.3 KB

Line
1	//
2	// Cforall Version 1.0.0 Copyright (C) 2017 University of Waterloo
3	//
4	// The contents of this file are covered under the licence agreement in the
5	// file "LICENCE" distributed with Cforall.
6	//
7	// heap.cfa --
8	//
9	// Author : Peter A. Buhr
10	// Created On : Tue Dec 19 21:58:35 2017
11	// Last Modified By : Peter A. Buhr
12	// Last Modified On : Sat Sep 30 17:31:15 2023
13	// Update Count : 1617
14	//
15
16	#include <stdio.h>
17	#include <string.h> // memset, memcpy
18	#include <limits.h> // ULONG_MAX
19	#include <errno.h> // errno, ENOMEM, EINVAL
20	#include <unistd.h> // STDERR_FILENO, sbrk, sysconf, write
21	#include <sys/mman.h> // mmap, munmap
22	extern "C" {
23	#include <sys/sysinfo.h> // get_nprocs
24	} // extern "C"
25
26	#include "heap.hfa"
27	#include "bits/align.hfa" // libAlign
28	#include "bits/defs.hfa" // likely, unlikely
29	#include "concurrency/kernel/fwd.hfa" // __POLL_PREEMPTION
30	#include "startup.hfa" // STARTUP_PRIORITY_MEMORY
31	#include "math.hfa" // ceiling, min
32	#include "bitmanip.hfa" // is_pow2, ceiling2
33
34	// supported mallopt options
35	#ifndef M_MMAP_THRESHOLD
36	#define M_MMAP_THRESHOLD (-1)
37	#endif // M_MMAP_THRESHOLD
38
39	#ifndef M_TOP_PAD
40	#define M_TOP_PAD (-2)
41	#endif // M_TOP_PAD
42
43	#define FASTLOOKUP // use O(1) table lookup from allocation size to bucket size
44	#define OWNERSHIP // return freed memory to owner thread
45	#define RETURNSPIN // toggle spinlock / lockfree queue
46	#if ! defined( OWNERSHIP ) && defined( RETURNSPIN )
47	#warning "RETURNSPIN is ignored without OWNERSHIP; suggest commenting out RETURNSPIN"
48	#endif // ! OWNERSHIP && RETURNSPIN
49
50	#define CACHE_ALIGN 64
51	#define CALIGN __attribute__(( aligned(CACHE_ALIGN) ))
52
53	#define TLSMODEL __attribute__(( tls_model("initial-exec") ))
54
55	#define __STATISTICS__
56
57	enum {
58	// The default extension heap amount in units of bytes. When the current heap reaches the brk address, the brk
59	// address is extended by the extension amount.
60	__CFA_DEFAULT_HEAP_EXPANSION__ = 10 * 1024 * 1024,
61
62	// The mmap crossover point during allocation. Allocations less than this amount are allocated from buckets; values
63	// greater than or equal to this value are mmap from the operating system.
64	__CFA_DEFAULT_MMAP_START__ = 512 * 1024 + 1,
65
66	// The default unfreed storage amount in units of bytes. When the uC++ program ends it subtracts this amount from
67	// the malloc/free counter to adjust for storage the program does not free.
68	__CFA_DEFAULT_HEAP_UNFREED__ = 0
69	}; // enum
70
71
72	//####################### Heap Trace/Print ####################
73
74
75	static bool traceHeap = false;
76
77	inline bool traceHeap() libcfa_public { return traceHeap; }
78
79	bool traceHeapOn() libcfa_public {
80	bool temp = traceHeap;
81	traceHeap = true;
82	return temp;
83	} // traceHeapOn
84
85	bool traceHeapOff() libcfa_public {
86	bool temp = traceHeap;
87	traceHeap = false;
88	return temp;
89	} // traceHeapOff
90
91	bool traceHeapTerm() libcfa_public { return false; }
92
93
94	static bool prtFree = false;
95
96	bool prtFree() {
97	return prtFree;
98	} // prtFree
99
100	bool prtFreeOn() {
101	bool temp = prtFree;
102	prtFree = true;
103	return temp;
104	} // prtFreeOn
105
106	bool prtFreeOff() {
107	bool temp = prtFree;
108	prtFree = false;
109	return temp;
110	} // prtFreeOff
111
112
113	//######################### Helpers #########################
114
115
116	// generic Bsearchl does not inline, so substitute with hand-coded binary-search.
117	inline __attribute__((always_inline))
118	static size_t Bsearchl( unsigned int key, const unsigned int vals[], size_t dim ) {
119	size_t l = 0, m, h = dim;
120	while ( l < h ) {
121	m = (l + h) / 2;
122	if ( (unsigned int &)(vals[m]) < key ) { // cast away const
123	l = m + 1;
124	} else {
125	h = m;
126	} // if
127	} // while
128	return l;
129	} // Bsearchl
130
131
132	// pause to prevent excess processor bus usage
133	#if defined( __i386 ) \|\| defined( __x86_64 )
134	#define Pause() __asm__ __volatile__ ( "pause" : : : )
135	#elif defined(__ARM_ARCH)
136	#define Pause() __asm__ __volatile__ ( "YIELD" : : : )
137	#else
138	#error unsupported architecture
139	#endif
140
141	typedef volatile uintptr_t SpinLock_t;
142
143	static inline __attribute__((always_inline)) void lock( volatile SpinLock_t & slock ) {
144	enum { SPIN_START = 4, SPIN_END = 64 * 1024, };
145	unsigned int spin = SPIN_START;
146
147	for ( unsigned int i = 1;; i += 1 ) {
148	if ( slock == 0 && __atomic_test_and_set( &slock, __ATOMIC_ACQUIRE ) == 0 ) break; // Fence
149	for ( volatile unsigned int s = 0; s < spin; s += 1 ) Pause(); // exponential spin
150	spin += spin; // powers of 2
151	//if ( i % 64 == 0 ) spin += spin; // slowly increase by powers of 2
152	if ( spin > SPIN_END ) spin = SPIN_END; // cap spinning
153	} // for
154	} // spin_lock
155
156	static inline __attribute__((always_inline)) void unlock( volatile SpinLock_t & slock ) {
157	__atomic_clear( &slock, __ATOMIC_RELEASE ); // Fence
158	} // spin_unlock
159
160
161	//####################### Heap Statistics ####################
162
163
164	#ifdef __STATISTICS__
165	enum { CntTriples = 12 }; // number of counter triples
166	enum { MALLOC, AALLOC, CALLOC, MEMALIGN, AMEMALIGN, CMEMALIGN, RESIZE, REALLOC, FREE };
167
168	struct StatsOverlay { // overlay for iteration
169	unsigned int calls, calls_0;
170	unsigned long long int request, alloc;
171	};
172
173	// Heap statistics counters.
174	union HeapStatistics {
175	struct { // minimum qualification
176	unsigned int malloc_calls, malloc_0_calls;
177	unsigned long long int malloc_storage_request, malloc_storage_alloc;
178	unsigned int aalloc_calls, aalloc_0_calls;
179	unsigned long long int aalloc_storage_request, aalloc_storage_alloc;
180	unsigned int calloc_calls, calloc_0_calls;
181	unsigned long long int calloc_storage_request, calloc_storage_alloc;
182	unsigned int memalign_calls, memalign_0_calls;
183	unsigned long long int memalign_storage_request, memalign_storage_alloc;
184	unsigned int amemalign_calls, amemalign_0_calls;
185	unsigned long long int amemalign_storage_request, amemalign_storage_alloc;
186	unsigned int cmemalign_calls, cmemalign_0_calls;
187	unsigned long long int cmemalign_storage_request, cmemalign_storage_alloc;
188	unsigned int resize_calls, resize_0_calls;
189	unsigned long long int resize_storage_request, resize_storage_alloc;
190	unsigned int realloc_calls, realloc_0_calls;
191	unsigned long long int realloc_storage_request, realloc_storage_alloc;
192	unsigned int free_calls, free_null_calls;
193	unsigned long long int free_storage_request, free_storage_alloc;
194	unsigned int return_pulls, return_pushes;
195	unsigned long long int return_storage_request, return_storage_alloc;
196	unsigned int mmap_calls, mmap_0_calls; // no zero calls
197	unsigned long long int mmap_storage_request, mmap_storage_alloc;
198	unsigned int munmap_calls, munmap_0_calls; // no zero calls
199	unsigned long long int munmap_storage_request, munmap_storage_alloc;
200	};
201	struct StatsOverlay counters[CntTriples]; // overlay for iteration
202	}; // HeapStatistics
203
204	static_assert( sizeof(HeapStatistics) == CntTriples * sizeof(StatsOverlay),
205	"Heap statistics counter-triplets does not match with array size" );
206
207	static void HeapStatisticsCtor( HeapStatistics & stats ) {
208	memset( &stats, '\0', sizeof(stats) ); // very fast
209	// for ( unsigned int i = 0; i < CntTriples; i += 1 ) {
210	// stats.counters[i].calls = stats.counters[i].calls_0 = stats.counters[i].request = stats.counters[i].alloc = 0;
211	// } // for
212	} // HeapStatisticsCtor
213
214	static HeapStatistics & ?+=?( HeapStatistics & lhs, const HeapStatistics & rhs ) {
215	for ( unsigned int i = 0; i < CntTriples; i += 1 ) {
216	lhs.counters[i].calls += rhs.counters[i].calls;
217	lhs.counters[i].calls_0 += rhs.counters[i].calls_0;
218	lhs.counters[i].request += rhs.counters[i].request;
219	lhs.counters[i].alloc += rhs.counters[i].alloc;
220	} // for
221	return lhs;
222	} // ?+=?
223	#endif // __STATISTICS__
224
225
226	// Recursive definitions: HeapManager needs size of bucket array and bucket area needs sizeof HeapManager storage.
227	// Break recursion by hardcoding number of buckets and statically checking number is correct after bucket array defined.
228	enum { NoBucketSizes = 91 }; // number of buckets sizes
229
230	struct Heap {
231	struct Storage {
232	struct Header { // header
233	union Kind {
234	struct RealHeader {
235	union {
236	struct { // 4-byte word => 8-byte header, 8-byte word => 16-byte header
237	union {
238	// 2nd low-order bit => zero filled, 3rd low-order bit => mmapped
239	// FreeHeader * home; // allocated block points back to home locations (must overlay alignment)
240	void * home; // allocated block points back to home locations (must overlay alignment)
241	size_t blockSize; // size for munmap (must overlay alignment)
242	Storage * next; // freed block points to next freed block of same size
243	};
244	size_t size; // allocation size in bytes
245	};
246	};
247	} real; // RealHeader
248
249	struct FakeHeader {
250	uintptr_t alignment; // 1st low-order bit => fake header & alignment
251	uintptr_t offset;
252	} fake; // FakeHeader
253	} kind; // Kind
254	} header; // Header
255
256	char pad[libAlign() - sizeof( Header )];
257	char data[0]; // storage
258	}; // Storage
259
260	static_assert( libAlign() >= sizeof( Storage ), "minimum alignment < sizeof( Storage )" );
261
262	struct CALIGN FreeHeader {
263	size_t blockSize CALIGN; // size of allocations on this list
264	#ifdef OWNERSHIP
265	#ifdef RETURNSPIN
266	SpinLock_t returnLock;
267	#endif // RETURNSPIN
268	Storage * returnList; // other thread return list
269	#endif // OWNERSHIP
270
271	Storage * freeList; // thread free list
272	Heap * homeManager; // heap owner (free storage to bucket, from bucket to heap)
273	}; // FreeHeader
274
275	FreeHeader freeLists[NoBucketSizes]; // buckets for different allocation sizes
276	void * heapBuffer; // start of free storage in buffer
277	size_t heapReserve; // amount of remaining free storage in buffer
278
279	#if defined( __STATISTICS__ ) \|\| defined( __CFA_DEBUG__ )
280	Heap * nextHeapManager; // intrusive link of existing heaps; traversed to collect statistics or check unfreed storage
281	#endif // __STATISTICS__ \|\| __CFA_DEBUG__
282	Heap * nextFreeHeapManager; // intrusive link of free heaps from terminated threads; reused by new threads
283
284	#ifdef __CFA_DEBUG__
285	ptrdiff_t allocUnfreed; // running total of allocations minus frees; can be negative
286	#endif // __CFA_DEBUG__
287
288	#ifdef __STATISTICS__
289	HeapStatistics stats; // local statistic table for this heap
290	#endif // __STATISTICS__
291	}; // Heap
292
293
294	struct HeapMaster {
295	SpinLock_t extLock; // protects allocation-buffer extension
296	SpinLock_t mgrLock; // protects freeHeapManagersList, heapManagersList, heapManagersStorage, heapManagersStorageEnd
297
298	void * heapBegin; // start of heap
299	void * heapEnd; // logical end of heap
300	size_t heapRemaining; // amount of storage not allocated in the current chunk
301	size_t pageSize; // architecture pagesize
302	size_t heapExpand; // sbrk advance
303	size_t mmapStart; // cross over point for mmap
304	unsigned int maxBucketsUsed; // maximum number of buckets in use
305
306	Heap * heapManagersList; // heap-list head
307	Heap * freeHeapManagersList; // free-list head
308
309	// Heap superblocks are not linked; heaps in superblocks are linked via intrusive links.
310	Heap * heapManagersStorage; // next heap to use in heap superblock
311	Heap * heapManagersStorageEnd; // logical heap outside of superblock's end
312
313	#ifdef __STATISTICS__
314	HeapStatistics stats; // global stats for thread-local heaps to add there counters when exiting
315	unsigned long int threads_started, threads_exited; // counts threads that have started and exited
316	unsigned long int reused_heap, new_heap; // counts reusability of heaps
317	unsigned int sbrk_calls;
318	unsigned long long int sbrk_storage;
319	int stats_fd;
320	#endif // __STATISTICS__
321	}; // HeapMaster
322
323
324	#ifdef FASTLOOKUP
325	enum { LookupSizes = 65_536 + sizeof(Heap.Storage) }; // number of fast lookup sizes
326	static unsigned char lookup[LookupSizes]; // O(1) lookup for small sizes
327	#endif // FASTLOOKUP
328
329	static volatile bool heapMasterBootFlag = false; // trigger for first heap
330	static HeapMaster heapMaster @= {}; // program global
331
332	static void heapMasterCtor();
333	static void heapMasterDtor();
334	static Heap * getHeap();
335
336
337	// Size of array must harmonize with NoBucketSizes and individual bucket sizes must be multiple of 16.
338	// Smaller multiples of 16 and powers of 2 are common allocation sizes, so make them generate the minimum required bucket size.
339	// malloc(0) returns 0p, so no bucket is necessary for 0 bytes returning an address that can be freed.
340	static const unsigned int bucketSizes[] @= { // different bucket sizes
341	16 + sizeof(Heap.Storage), 32 + sizeof(Heap.Storage), 48 + sizeof(Heap.Storage), 64 + sizeof(Heap.Storage), // 4
342	96 + sizeof(Heap.Storage), 112 + sizeof(Heap.Storage), 128 + sizeof(Heap.Storage), // 3
343	160, 192, 224, 256 + sizeof(Heap.Storage), // 4
344	320, 384, 448, 512 + sizeof(Heap.Storage), // 4
345	640, 768, 896, 1_024 + sizeof(Heap.Storage), // 4
346	1_536, 2_048 + sizeof(Heap.Storage), // 2
347	2_560, 3_072, 3_584, 4_096 + sizeof(Heap.Storage), // 4
348	6_144, 8_192 + sizeof(Heap.Storage), // 2
349	9_216, 10_240, 11_264, 12_288, 13_312, 14_336, 15_360, 16_384 + sizeof(Heap.Storage), // 8
350	18_432, 20_480, 22_528, 24_576, 26_624, 28_672, 30_720, 32_768 + sizeof(Heap.Storage), // 8
351	36_864, 40_960, 45_056, 49_152, 53_248, 57_344, 61_440, 65_536 + sizeof(Heap.Storage), // 8
352	73_728, 81_920, 90_112, 98_304, 106_496, 114_688, 122_880, 131_072 + sizeof(Heap.Storage), // 8
353	147_456, 163_840, 180_224, 196_608, 212_992, 229_376, 245_760, 262_144 + sizeof(Heap.Storage), // 8
354	294_912, 327_680, 360_448, 393_216, 425_984, 458_752, 491_520, 524_288 + sizeof(Heap.Storage), // 8
355	655_360, 786_432, 917_504, 1_048_576 + sizeof(Heap.Storage), // 4
356	1_179_648, 1_310_720, 1_441_792, 1_572_864, 1_703_936, 1_835_008, 1_966_080, 2_097_152 + sizeof(Heap.Storage), // 8
357	2_621_440, 3_145_728, 3_670_016, 4_194_304 + sizeof(Heap.Storage), // 4
358	};
359
360	static_assert( NoBucketSizes == sizeof(bucketSizes) / sizeof(bucketSizes[0] ), "size of bucket array wrong" );
361
362
363	// extern visibility, used by runtime kernel
364	libcfa_public size_t __page_size; // architecture pagesize
365	libcfa_public int __map_prot; // common mmap/mprotect protection
366
367
368	// Thread-local storage is allocated lazily when the storage is accessed.
369	static __thread size_t PAD1 CALIGN TLSMODEL __attribute__(( unused )); // protect false sharing
370	static __thread Heap * heapManager CALIGN TLSMODEL;
371	static __thread bool heapManagerBootFlag CALIGN TLSMODEL = false;
372	static __thread size_t PAD2 CALIGN TLSMODEL __attribute__(( unused )); // protect further false sharing
373
374
375	// declare helper functions for HeapMaster
376	void noMemory(); // forward, called by "builtin_new" when malloc returns 0
377
378
379	void heapMasterCtor() with( heapMaster ) {
380	// Singleton pattern to initialize heap master
381
382	verify( bucketSizes[0] == (16 + sizeof(Heap.Storage)) );
383
384	__page_size = sysconf( _SC_PAGESIZE );
385	__map_prot = PROT_READ \| PROT_WRITE \| PROT_EXEC;
386
387	extLock = 0;
388	mgrLock = 0;
389
390	char * end = (char *)sbrk( 0 );
391	heapBegin = heapEnd = sbrk( (char *)ceiling2( (long unsigned int)end, libAlign() ) - end ); // move start of heap to multiple of alignment
392	heapRemaining = 0;
393	heapExpand = malloc_expansion();
394	mmapStart = malloc_mmap_start();
395
396	// find the closest bucket size less than or equal to the mmapStart size
397	maxBucketsUsed = Bsearchl( mmapStart, bucketSizes, NoBucketSizes ); // binary search
398
399	verify( (mmapStart >= pageSize) && (bucketSizes[NoBucketSizes - 1] >= mmapStart) );
400	verify( maxBucketsUsed < NoBucketSizes ); // subscript failure ?
401	verify( mmapStart <= bucketSizes[maxBucketsUsed] ); // search failure ?
402
403	heapManagersList = 0p;
404	freeHeapManagersList = 0p;
405
406	heapManagersStorage = 0p;
407	heapManagersStorageEnd = 0p;
408
409	#ifdef __STATISTICS__
410	HeapStatisticsCtor( stats ); // clear statistic counters
411	threads_started = threads_exited = 0;
412	reused_heap = new_heap = 0;
413	sbrk_calls = sbrk_storage = 0;
414	stats_fd = STDERR_FILENO;
415	#endif // __STATISTICS__
416
417	#ifdef FASTLOOKUP
418	for ( unsigned int i = 0, idx = 0; i < LookupSizes; i += 1 ) {
419	if ( i > bucketSizes[idx] ) idx += 1;
420	lookup[i] = idx;
421	verify( i <= bucketSizes[idx] );
422	verify( (i <= 32 && idx == 0) \|\| (i > bucketSizes[idx - 1]) );
423	} // for
424	#endif // FASTLOOKUP
425
426	heapMasterBootFlag = true;
427	} // heapMasterCtor
428
429
430	#define NO_MEMORY_MSG "** Error ** insufficient heap memory available to allocate %zd new bytes."
431
432	Heap * getHeap() with( heapMaster ) {
433	Heap * heap;
434	if ( freeHeapManagersList ) { // free heap for reused ?
435	heap = freeHeapManagersList;
436	freeHeapManagersList = heap->nextFreeHeapManager;
437
438	#ifdef __STATISTICS__
439	reused_heap += 1;
440	#endif // __STATISTICS__
441	} else { // free heap not found, create new
442	// Heap size is about 12K, FreeHeader (128 bytes because of cache alignment) * NoBucketSizes (91) => 128 heaps *
443	// 12K ~= 120K byte superblock. Where 128-heap superblock handles a medium sized multi-processor server.
444	size_t remaining = heapManagersStorageEnd - heapManagersStorage; // remaining free heaps in superblock
445	if ( ! heapManagersStorage \|\| remaining == 0 ) {
446	// Each block of heaps is a multiple of the number of cores on the computer.
447	int HeapDim = get_nprocs(); // get_nprocs_conf does not work
448	size_t size = HeapDim * sizeof( Heap );
449
450	heapManagersStorage = (Heap *)mmap( 0, size, PROT_READ \| PROT_WRITE, MAP_PRIVATE \| MAP_ANONYMOUS, -1, 0 );
451	if ( unlikely( heapManagersStorage == (Heap *)MAP_FAILED ) ) { // failed ?
452	if ( errno == ENOMEM ) abort( NO_MEMORY_MSG, size ); // no memory
453	// Do not call strerror( errno ) as it may call malloc.
454	abort( "** Error ** attempt to allocate block of heaps of size %zu bytes and mmap failed with errno %d.", size, errno );
455	} // if
456	heapManagersStorageEnd = &heapManagersStorage[HeapDim]; // outside array
457	} // if
458
459	heap = heapManagersStorage;
460	heapManagersStorage = heapManagersStorage + 1; // bump next heap
461
462	#if defined( __STATISTICS__ ) \|\| defined( __CFA_DEBUG__ )
463	heap->nextHeapManager = heapManagersList;
464	#endif // __STATISTICS__ \|\| __CFA_DEBUG__
465	heapManagersList = heap;
466
467	#ifdef __STATISTICS__
468	new_heap += 1;
469	#endif // __STATISTICS__
470
471	with( *heap ) {
472	for ( unsigned int j = 0; j < NoBucketSizes; j += 1 ) { // initialize free lists
473	#ifdef OWNERSHIP
474	#ifdef RETURNSPIN
475	freeLists[j].returnLock = 0;
476	freeLists[j].returnList = 0p;
477	#endif // RETURNSPIN
478	#endif // OWNERSHIP
479
480	freeLists[j].freeList = 0p;
481	freeLists[j].homeManager = heap;
482	freeLists[j].blockSize = bucketSizes[j];
483	} // for
484
485	heapBuffer = 0p;
486	heapReserve = 0;
487	nextFreeHeapManager = 0p;
488	#ifdef __CFA_DEBUG__
489	allocUnfreed = 0;
490	#endif // __CFA_DEBUG__
491	heapManagerBootFlag = true;
492	} // with
493	} // if
494
495	return heap;
496	} // getHeap
497
498
499	void heapManagerCtor() libcfa_public {
500	if ( unlikely( ! heapMasterBootFlag ) ) heapMasterCtor();
501
502	lock( heapMaster.mgrLock ); // protect heapMaster counters
503
504	assert( ! heapManagerBootFlag );
505
506	// get storage for heap manager
507
508	heapManager = getHeap();
509
510	#ifdef __STATISTICS__
511	HeapStatisticsCtor( heapManager->stats ); // heap local
512	heapMaster.threads_started += 1;
513	#endif // __STATISTICS__
514
515	unlock( heapMaster.mgrLock );
516	} // heapManagerCtor
517
518
519	void heapManagerDtor() libcfa_public {
520	if ( unlikely( ! heapManagerBootFlag ) ) return; // thread never used ?
521
522	lock( heapMaster.mgrLock );
523
524	// place heap on list of free heaps for reusability
525	heapManager->nextFreeHeapManager = heapMaster.freeHeapManagersList;
526	heapMaster.freeHeapManagersList = heapManager;
527
528	#ifdef __STATISTICS__
529	heapMaster.stats += heapManager->stats; // retain this heap's statistics
530	heapMaster.threads_exited += 1;
531	#endif // __STATISTICS__
532
533	// Do not set heapManager to NULL because it is used after Cforall is shutdown but before the program shuts down.
534
535	heapManagerBootFlag = false;
536	unlock( heapMaster.mgrLock );
537	} // heapManagerDtor
538
539
540	//####################### Memory Allocation Routines Helpers ####################
541
542
543	extern int cfa_main_returned; // from interpose.cfa
544	extern "C" {
545	void memory_startup( void ) { // singleton => called once at start of program
546	if ( ! heapMasterBootFlag ) heapManagerCtor(); // sanity check
547	} // memory_startup
548
549	void memory_shutdown( void ) {
550	heapManagerDtor();
551	} // memory_shutdown
552
553	void heapAppStart() { // called by __cfaabi_appready_startup
554	verify( heapManager );
555	#ifdef __CFA_DEBUG__
556	heapManager->allocUnfreed = 0; // clear prior allocation counts
557	#endif // __CFA_DEBUG__
558
559	#ifdef __STATISTICS__
560	HeapStatisticsCtor( heapManager->stats ); // clear prior statistic counters
561	#endif // __STATISTICS__
562	} // heapAppStart
563
564	void heapAppStop() { // called by __cfaabi_appready_startdown
565	fclose( stdin ); fclose( stdout ); // free buffer storage
566	if ( ! cfa_main_returned ) return; // do not check unfreed storage if exit called
567
568	#ifdef __STATISTICS__
569	if ( getenv( "CFA_MALLOC_STATS" ) ) { // check for external printing
570	malloc_stats();
571	} // if
572	#endif // __STATISTICS__
573
574	#ifdef __CFA_DEBUG__
575	// allocUnfreed is set to 0 when a heap is created and it accumulates any unfreed storage during its multiple thread
576	// usages. At the end, add up each heap allocUnfreed value across all heaps to get the total unfreed storage.
577	ptrdiff_t allocUnfreed = 0;
578	for ( Heap * heap = heapMaster.heapManagersList; heap; heap = heap->nextHeapManager ) {
579	allocUnfreed += heap->allocUnfreed;
580	} // for
581
582	allocUnfreed -= malloc_unfreed(); // subtract any user specified unfreed storage
583	if ( allocUnfreed > 0 ) {
584	// DO NOT USE STREAMS AS THEY MAY BE UNAVAILABLE AT THIS POINT.
585	char helpText[512];
586	__cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
587	"CFA warning (UNIX pid:%ld) : program terminating with %td(%#tx) bytes of storage allocated but not freed.\n"
588	"Possible cause is unfreed storage allocated by the program or system/library routines called from the program.\n",
589	(long int)getpid(), allocUnfreed, allocUnfreed ); // always print the UNIX pid
590	} // if
591	#endif // __CFA_DEBUG__
592	} // heapAppStop
593	} // extern "C"
594
595
596	#ifdef __STATISTICS__
597	static HeapStatistics stats; // zero filled
598
599	#define prtFmt \
600	"\nHeap statistics: (storage request / allocation)\n" \
601	" malloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
602	" aalloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
603	" calloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
604	" memalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
605	" amemalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
606	" cmemalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
607	" resize >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
608	" realloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
609	" free !null calls %'u; null calls %'u; storage %'llu / %'llu bytes\n" \
610	" return pulls %'u; pushes %'u; storage %'llu / %'llu bytes\n" \
611	" sbrk calls %'u; storage %'llu bytes\n" \
612	" mmap calls %'u; storage %'llu / %'llu bytes\n" \
613	" munmap calls %'u; storage %'llu / %'llu bytes\n" \
614	" threads started %'lu; exited %'lu\n" \
615	" heaps new %'lu; reused %'lu\n"
616
617	// Use "write" because streams may be shutdown when calls are made.
618	static int printStats( HeapStatistics & stats ) with( heapMaster, stats ) { // see malloc_stats
619	char helpText[sizeof(prtFmt) + 1024]; // space for message and values
620	return __cfaabi_bits_print_buffer( stats_fd, helpText, sizeof(helpText), prtFmt,
621	malloc_calls, malloc_0_calls, malloc_storage_request, malloc_storage_alloc,
622	aalloc_calls, aalloc_0_calls, aalloc_storage_request, aalloc_storage_alloc,
623	calloc_calls, calloc_0_calls, calloc_storage_request, calloc_storage_alloc,
624	memalign_calls, memalign_0_calls, memalign_storage_request, memalign_storage_alloc,
625	amemalign_calls, amemalign_0_calls, amemalign_storage_request, amemalign_storage_alloc,
626	cmemalign_calls, cmemalign_0_calls, cmemalign_storage_request, cmemalign_storage_alloc,
627	resize_calls, resize_0_calls, resize_storage_request, resize_storage_alloc,
628	realloc_calls, realloc_0_calls, realloc_storage_request, realloc_storage_alloc,
629	free_calls, free_null_calls, free_storage_request, free_storage_alloc,
630	return_pulls, return_pushes, return_storage_request, return_storage_alloc,
631	sbrk_calls, sbrk_storage,
632	mmap_calls, mmap_storage_request, mmap_storage_alloc,
633	munmap_calls, munmap_storage_request, munmap_storage_alloc,
634	threads_started, threads_exited,
635	new_heap, reused_heap
636	);
637	} // printStats
638
639	#define prtFmtXML \
640	"<malloc version=\"1\">\n" \
641	"<heap nr=\"0\">\n" \
642	"<sizes>\n" \
643	"</sizes>\n" \
644	"<total type=\"malloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
645	"<total type=\"aalloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
646	"<total type=\"calloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
647	"<total type=\"memalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
648	"<total type=\"amemalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
649	"<total type=\"cmemalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
650	"<total type=\"resize\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
651	"<total type=\"realloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
652	"<total type=\"free\" !null=\"%'u;\" 0 null=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
653	"<total type=\"return\" pulls=\"%'u;\" 0 pushes=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
654	"<total type=\"sbrk\" count=\"%'u;\" size=\"%'llu\"/> bytes\n" \
655	"<total type=\"mmap\" count=\"%'u;\" size=\"%'llu / %'llu\" / > bytes\n" \
656	"<total type=\"munmap\" count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
657	"<total type=\"threads\" started=\"%'lu;\" exited=\"%'lu\"/>\n" \
658	"<total type=\"heaps\" new=\"%'lu;\" reused=\"%'lu\"/>\n" \
659	"</malloc>"
660
661	static int printStatsXML( HeapStatistics & stats, FILE * stream ) with( heapMaster, stats ) { // see malloc_info
662	char helpText[sizeof(prtFmtXML) + 1024]; // space for message and values
663	return __cfaabi_bits_print_buffer( fileno( stream ), helpText, sizeof(helpText), prtFmtXML,
664	malloc_calls, malloc_0_calls, malloc_storage_request, malloc_storage_alloc,
665	aalloc_calls, aalloc_0_calls, aalloc_storage_request, aalloc_storage_alloc,
666	calloc_calls, calloc_0_calls, calloc_storage_request, calloc_storage_alloc,
667	memalign_calls, memalign_0_calls, memalign_storage_request, memalign_storage_alloc,
668	amemalign_calls, amemalign_0_calls, amemalign_storage_request, amemalign_storage_alloc,
669	cmemalign_calls, cmemalign_0_calls, cmemalign_storage_request, cmemalign_storage_alloc,
670	resize_calls, resize_0_calls, resize_storage_request, resize_storage_alloc,
671	realloc_calls, realloc_0_calls, realloc_storage_request, realloc_storage_alloc,
672	free_calls, free_null_calls, free_storage_request, free_storage_alloc,
673	return_pulls, return_pushes, return_storage_request, return_storage_alloc,
674	sbrk_calls, sbrk_storage,
675	mmap_calls, mmap_storage_request, mmap_storage_alloc,
676	munmap_calls, munmap_storage_request, munmap_storage_alloc,
677	threads_started, threads_exited,
678	new_heap, reused_heap
679	);
680	} // printStatsXML
681
682	static HeapStatistics & collectStats( HeapStatistics & stats ) with( heapMaster ) {
683	lock( heapMaster.mgrLock );
684
685	stats += heapMaster.stats;
686	for ( Heap * heap = heapManagersList; heap; heap = heap->nextHeapManager ) {
687	stats += heap->stats;
688	} // for
689
690	unlock( heapMaster.mgrLock );
691	return stats;
692	} // collectStats
693
694	static inline void clearStats() {
695	lock( heapMaster.mgrLock );
696
697	// Zero the heap master and all active thread heaps.
698	HeapStatisticsCtor( heapMaster.stats );
699	for ( Heap * heap = heapMaster.heapManagersList; heap; heap = heap->nextHeapManager ) {
700	HeapStatisticsCtor( heap->stats );
701	} // for
702
703	unlock( heapMaster.mgrLock );
704	} // clearStats
705	#endif // __STATISTICS__
706
707
708	static bool setMmapStart( size_t value ) with( heapMaster ) { // true => mmapped, false => sbrk
709	if ( value < __page_size \|\| bucketSizes[NoBucketSizes - 1] < value ) return false;
710	mmapStart = value; // set global
711
712	// find the closest bucket size less than or equal to the mmapStart size
713	maxBucketsUsed = Bsearchl( mmapStart, bucketSizes, NoBucketSizes ); // binary search
714
715	verify( maxBucketsUsed < NoBucketSizes ); // subscript failure ?
716	verify( mmapStart <= bucketSizes[maxBucketsUsed] ); // search failure ?
717	return true;
718	} // setMmapStart
719
720
721	// <-------+----------------------------------------------------> bsize (bucket size)
722	// \|header \|addr
723	//==================================================================================
724	// align/offset \|
725	// <-----------------<------------+-----------------------------> bsize (bucket size)
726	// \|fake-header \| addr
727	#define HeaderAddr( addr ) ((Heap.Storage.Header )( (char )addr - sizeof(Heap.Storage) ))
728	#define RealHeader( header ) ((Heap.Storage.Header )((char )header - header->kind.fake.offset))
729
730	// <-------<<--------------------- dsize ---------------------->> bsize (bucket size)
731	// \|header \|addr
732	//==================================================================================
733	// align/offset \|
734	// <------------------------------<<---------- dsize --------->>> bsize (bucket size)
735	// \|fake-header \|addr
736	#define DataStorage( bsize, addr, header ) (bsize - ( (char )addr - (char )header ))
737
738
739	inline __attribute__((always_inline))
740	static void checkAlign( size_t alignment ) {
741	if ( unlikely( alignment < libAlign() \|\| ! is_pow2( alignment ) ) ) {
742	abort( "** Error ** alignment %zu for memory allocation is less than %d and/or not a power of 2.", alignment, libAlign() );
743	} // if
744	} // checkAlign
745
746
747	inline __attribute__((always_inline))
748	static void checkHeader( bool check, const char name[], void * addr ) {
749	if ( unlikely( check ) ) { // bad address ?
750	abort( "** Error ** attempt to %s storage %p with address outside the heap.\n"
751	"Possible cause is duplicate free on same block or overwriting of memory.",
752	name, addr );
753	} // if
754	} // checkHeader
755
756
757	// Manipulate sticky bits stored in unused 3 low-order bits of an address.
758	// bit0 => alignment => fake header
759	// bit1 => zero filled (calloc)
760	// bit2 => mapped allocation versus sbrk
761	#define StickyBits( header ) (((header)->kind.real.blockSize & 0x7))
762	#define ClearStickyBits( addr ) (typeof(addr))((uintptr_t)(addr) & ~7)
763	#define MarkAlignmentBit( align ) ((align) \| 1)
764	#define AlignmentBit( header ) ((((header)->kind.fake.alignment) & 1))
765	#define ClearAlignmentBit( header ) (((header)->kind.fake.alignment) & ~1)
766	#define ZeroFillBit( header ) ((((header)->kind.real.blockSize) & 2))
767	#define ClearZeroFillBit( header ) ((((header)->kind.real.blockSize) &= ~2))
768	#define MarkZeroFilledBit( header ) ((header)->kind.real.blockSize \|= 2)
769	#define MmappedBit( header ) ((((header)->kind.real.blockSize) & 4))
770	#define MarkMmappedBit( size ) ((size) \| 4)
771
772
773	inline __attribute__((always_inline))
774	static void fakeHeader( Heap.Storage.Header *& header, size_t & alignment ) {
775	if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
776	alignment = ClearAlignmentBit( header ); // clear flag from value
777	#ifdef __CFA_DEBUG__
778	checkAlign( alignment ); // check alignment
779	#endif // __CFA_DEBUG__
780	header = RealHeader( header ); // backup from fake to real header
781	} else {
782	alignment = libAlign(); // => no fake header
783	} // if
784	} // fakeHeader
785
786
787	inline __attribute__((always_inline))
788	static bool headers( const char name[] __attribute__(( unused )), void * addr, Heap.Storage.Header *& header,
789	Heap.FreeHeader & freeHead, size_t & size, size_t & alignment ) with( heapMaster, heapManager ) {
790	header = HeaderAddr( addr );
791
792	#ifdef __CFA_DEBUG__
793	checkHeader( header < (Heap.Storage.Header *)heapBegin, name, addr ); // bad low address ?
794	#endif // __CFA_DEBUG__
795
796	if ( likely( ! StickyBits( header ) ) ) { // no sticky bits ?
797	freeHead = (Heap.FreeHeader *)(header->kind.real.home);
798	alignment = libAlign();
799	} else {
800	fakeHeader( header, alignment );
801	if ( unlikely( MmappedBit( header ) ) ) { // mmapped ?
802	verify( addr < heapBegin \|\| heapEnd < addr );
803	size = ClearStickyBits( header->kind.real.blockSize ); // mmap size
804	return true;
805	} // if
806
807	freeHead = (Heap.FreeHeader *)(ClearStickyBits( header->kind.real.home ));
808	} // if
809	size = freeHead->blockSize;
810
811	#ifdef __CFA_DEBUG__
812	checkHeader( header < (Heap.Storage.Header )heapBegin \|\| (Heap.Storage.Header )heapEnd < header, name, addr ); // bad address ? (offset could be + or -)
813
814	Heap * homeManager;
815	if ( unlikely( freeHead == 0p \|\| // freed and only free-list node => null link
816	// freed and link points at another free block not to a bucket in the bucket array.
817	(homeManager = freeHead->homeManager, freeHead < &homeManager->freeLists[0] \|\|
818	&homeManager->freeLists[NoBucketSizes] <= freeHead ) ) ) {
819	abort( "** Error ** attempt to %s storage %p with corrupted header.\n"
820	"Possible cause is duplicate free on same block or overwriting of header information.",
821	name, addr );
822	} // if
823	#endif // __CFA_DEBUG__
824
825	return false;
826	} // headers
827
828
829	static void * master_extend( size_t size ) with( heapMaster ) {
830	lock( extLock );
831
832	ptrdiff_t rem = heapRemaining - size;
833	if ( unlikely( rem < 0 ) ) { // negative ?
834	// If the size requested is bigger than the current remaining storage, increase the size of the heap.
835
836	size_t increase = ceiling2( size > heapExpand ? size : heapExpand, libAlign() );
837	if ( unlikely( sbrk( increase ) == (void *)-1 ) ) { // failed, no memory ?
838	unlock( extLock );
839	abort( NO_MEMORY_MSG, size ); // give up
840	} // if
841
842	// Make storage executable for thunks.
843	if ( mprotect( (char *)heapEnd + heapRemaining, increase, __map_prot ) ) {
844	unlock( extLock );
845	abort( "** Error ** attempt to make heap storage executable for thunks and mprotect failed with errno %d.", errno );
846	} // if
847
848	rem = heapRemaining + increase - size;
849
850	#ifdef __STATISTICS__
851	sbrk_calls += 1;
852	sbrk_storage += increase;
853	#endif // __STATISTICS__
854	} // if
855
856	Heap.Storage * block = (Heap.Storage *)heapEnd;
857	heapRemaining = rem;
858	heapEnd = (char *)heapEnd + size;
859
860	unlock( extLock );
861	return block;
862	} // master_extend
863
864
865	__attribute__(( noinline ))
866	static void * manager_extend( size_t size ) with( *heapManager ) {
867	ptrdiff_t rem = heapReserve - size;
868
869	if ( unlikely( rem < 0 ) ) { // negative ?
870	// If the size requested is bigger than the current remaining reserve, use the current reserve to populate
871	// smaller freeLists, and increase the reserve.
872
873	rem = heapReserve; // positive
874
875	if ( (unsigned int)rem >= bucketSizes[0] ) { // minimal size ? otherwise ignore
876	size_t bucket;
877	#ifdef FASTLOOKUP
878	if ( likely( rem < LookupSizes ) ) bucket = lookup[rem];
879	#endif // FASTLOOKUP
880	bucket = Bsearchl( rem, bucketSizes, heapMaster.maxBucketsUsed );
881	verify( 0 <= bucket && bucket <= heapMaster.maxBucketsUsed );
882	Heap.FreeHeader * freeHead = &(freeLists[bucket]);
883
884	// The remaining storage may not be bucket size, whereas all other allocations are. Round down to previous
885	// bucket size in this case.
886	if ( unlikely( freeHead->blockSize > (size_t)rem ) ) freeHead -= 1;
887	Heap.Storage * block = (Heap.Storage *)heapBuffer;
888
889	block->header.kind.real.next = freeHead->freeList; // push on stack
890	freeHead->freeList = block;
891	} // if
892
893	size_t increase = ceiling( size > ( heapMaster.heapExpand / 10 ) ? size : ( heapMaster.heapExpand / 10 ), libAlign() );
894	heapBuffer = master_extend( increase );
895	rem = increase - size;
896	} // if
897
898	Heap.Storage * block = (Heap.Storage *)heapBuffer;
899	heapReserve = rem;
900	heapBuffer = (char *)heapBuffer + size;
901
902	return block;
903	} // manager_extend
904
905
906	#define BOOT_HEAP_MANAGER \
907	if ( unlikely( ! heapMasterBootFlag ) ) { \
908	heapManagerCtor(); /* trigger for first heap */ \
909	} /* if */
910
911	#ifdef __STATISTICS__
912	#define STAT_NAME __counter
913	#define STAT_PARM , unsigned int STAT_NAME
914	#define STAT_ARG( name ) , name
915	#define STAT_0_CNT( counter ) stats.counters[counter].calls_0 += 1
916	#else
917	#define STAT_NAME
918	#define STAT_PARM
919	#define STAT_ARG( name )
920	#define STAT_0_CNT( counter )
921	#endif // __STATISTICS__
922
923	// Uncomment to get allocation addresses for a 0-sized allocation rather than a null pointer.
924	//#define __NONNULL_0_ALLOC__
925	#if ! defined( __NONNULL_0_ALLOC__ )
926	#define __NULL_0_ALLOC__ unlikely( size == 0 ) \|\| /* 0 BYTE ALLOCATION RETURNS NULL POINTER */
927	#else
928	#define __NULL_0_ALLOC__
929	#endif // __NONNULL_0_ALLOC__
930
931	#define PROLOG( counter, ... ) \
932	BOOT_HEAP_MANAGER; \
933	if ( \
934	__NULL_0_ALLOC__ \
935	unlikely( size > ULONG_MAX - sizeof(Heap.Storage) ) ) { /* error check */ \
936	STAT_0_CNT( counter ); \
937	__VA_ARGS__; \
938	return 0p; \
939	} /* if */
940
941
942	#define SCRUB_SIZE 1024lu
943	// Do not use '\xfe' for scrubbing because dereferencing an address composed of it causes a SIGSEGV without a valid IP
944	// pointer in the interrupt frame.
945	#define SCRUB '\xff'
946
947	static void * doMalloc( size_t size STAT_PARM ) libcfa_nopreempt with( *heapManager ) {
948	PROLOG( STAT_NAME );
949
950	verify( heapManager );
951	Heap.Storage * block; // pointer to new block of storage
952
953	// Look up size in the size list. Make sure the user request includes space for the header that must be allocated
954	// along with the block and is a multiple of the alignment size.
955	size_t tsize = size + sizeof(Heap.Storage);
956
957	#ifdef __STATISTICS__
958	stats.counters[STAT_NAME].calls += 1;
959	stats.counters[STAT_NAME].request += size;
960	#endif // __STATISTICS__
961
962	#ifdef __CFA_DEBUG__
963	allocUnfreed += size;
964	#endif // __CFA_DEBUG__
965
966	if ( likely( tsize < heapMaster.mmapStart ) ) { // small size => sbrk
967	size_t bucket;
968	#ifdef FASTLOOKUP
969	if ( likely( tsize < LookupSizes ) ) bucket = lookup[tsize];
970	else
971	#endif // FASTLOOKUP
972	bucket = Bsearchl( tsize, bucketSizes, heapMaster.maxBucketsUsed );
973	verify( 0 <= bucket && bucket <= heapMaster.maxBucketsUsed );
974	Heap.FreeHeader * freeHead = &freeLists[bucket];
975
976	verify( freeHead <= &freeLists[heapMaster.maxBucketsUsed] ); // subscripting error ?
977	verify( tsize <= freeHead->blockSize ); // search failure ?
978
979	tsize = freeHead->blockSize; // total space needed for request
980	#ifdef __STATISTICS__
981	stats.counters[STAT_NAME].alloc += tsize;
982	#endif // __STATISTICS__
983
984	block = freeHead->freeList; // remove node from stack
985	if ( unlikely( block == 0p ) ) { // no free block ?
986	// Freelist for this size is empty, so check return list (OWNERSHIP), or carve it out of the heap if there
987	// is enough left, or get some more heap storage and carve it off.
988	#ifdef OWNERSHIP
989	if ( unlikely( freeHead->returnList ) ) { // race, get next time if lose race
990	#ifdef RETURNSPIN
991	lock( freeHead->returnLock );
992	block = freeHead->returnList;
993	freeHead->returnList = 0p;
994	unlock( freeHead->returnLock );
995	#else
996	block = __atomic_exchange_n( &freeHead->returnList, 0p, __ATOMIC_SEQ_CST );
997	#endif // RETURNSPIN
998
999	verify( block );
1000	#ifdef __STATISTICS__
1001	stats.return_pulls += 1;
1002	#endif // __STATISTICS__
1003
1004	// OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
1005
1006	freeHead->freeList = block->header.kind.real.next; // merge returnList into freeHead
1007	} else {
1008	#endif // OWNERSHIP
1009	// Do not leave kernel thread as manager_extend accesses heapManager.
1010	disable_interrupts();
1011	block = (Heap.Storage *)manager_extend( tsize ); // mutual exclusion on call
1012	enable_interrupts( false );
1013
1014	// OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
1015
1016	#ifdef __CFA_DEBUG__
1017	// Scrub new memory so subsequent uninitialized usages might fail. Only scrub the first SCRUB_SIZE bytes.
1018	memset( block->data, SCRUB, min( SCRUB_SIZE, tsize - sizeof(Heap.Storage) ) );
1019	#endif // __CFA_DEBUG__
1020	#ifdef OWNERSHIP
1021	} // if
1022	#endif // OWNERSHIP
1023	} else {
1024	// Memory is scrubbed in doFree.
1025	freeHead->freeList = block->header.kind.real.next;
1026	} // if
1027
1028	block->header.kind.real.home = freeHead; // pointer back to free list of apropriate size
1029	} else { // large size => mmap
1030	if ( unlikely( size > ULONG_MAX - __page_size ) ) return 0p;
1031	tsize = ceiling2( tsize, __page_size ); // must be multiple of page size
1032
1033	#ifdef __STATISTICS__
1034	stats.counters[STAT_NAME].alloc += tsize;
1035	stats.mmap_calls += 1;
1036	stats.mmap_storage_request += size;
1037	stats.mmap_storage_alloc += tsize;
1038	#endif // __STATISTICS__
1039
1040	disable_interrupts();
1041	block = (Heap.Storage *)mmap( 0, tsize, __map_prot, MAP_PRIVATE \| MAP_ANONYMOUS, -1, 0 );
1042	enable_interrupts( false );
1043
1044	// OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
1045
1046	if ( unlikely( block == (Heap.Storage *)MAP_FAILED ) ) { // failed ?
1047	if ( errno == ENOMEM ) abort( NO_MEMORY_MSG, tsize ); // no memory
1048	// Do not call strerror( errno ) as it may call malloc.
1049	abort( "** Error ** attempt to allocate large object (> %zu) of size %zu bytes and mmap failed with errno %d.",
1050	size, heapMaster.mmapStart, errno );
1051	} // if
1052	block->header.kind.real.blockSize = MarkMmappedBit( tsize ); // storage size for munmap
1053
1054	#ifdef __CFA_DEBUG__
1055	// Scrub new memory so subsequent uninitialized usages might fail. Only scrub the first SCRUB_SIZE bytes. The
1056	// rest of the storage set to 0 by mmap.
1057	memset( block->data, SCRUB, min( SCRUB_SIZE, tsize - sizeof(Heap.Storage) ) );
1058	#endif // __CFA_DEBUG__
1059	} // if
1060
1061	block->header.kind.real.size = size; // store allocation size
1062	void * addr = &(block->data); // adjust off header to user bytes
1063	verify( ((uintptr_t)addr & (libAlign() - 1)) == 0 ); // minimum alignment ?
1064
1065	#ifdef __CFA_DEBUG__
1066	if ( traceHeap() ) {
1067	char helpText[64];
1068	__cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
1069	"%p = Malloc( %zu ) (allocated %zu)\n", addr, size, tsize ); // print debug/nodebug
1070	} // if
1071	#endif // __CFA_DEBUG__
1072
1073	// poll_interrupts(); // call rollforward
1074
1075	return addr;
1076	} // doMalloc
1077
1078
1079	static void doFree( void * addr ) libcfa_nopreempt with( *heapManager ) {
1080	verify( addr );
1081
1082	// detect free after thread-local storage destruction and use global stats in that case
1083
1084	Heap.Storage.Header * header;
1085	Heap.FreeHeader * freeHead;
1086	size_t size, alignment;
1087
1088	bool mapped = headers( "free", addr, header, freeHead, size, alignment );
1089	#if defined( __STATISTICS__ ) \|\| defined( __CFA_DEBUG__ )
1090	size_t rsize = header->kind.real.size; // optimization
1091	#endif // __STATISTICS__ \|\| __CFA_DEBUG__
1092
1093	#ifdef __STATISTICS__
1094	stats.free_storage_request += rsize;
1095	stats.free_storage_alloc += size;
1096	#endif // __STATISTICS__
1097
1098	#ifdef __CFA_DEBUG__
1099	allocUnfreed -= rsize;
1100	#endif // __CFA_DEBUG__
1101
1102	if ( unlikely( mapped ) ) { // mmapped ?
1103	#ifdef __STATISTICS__
1104	stats.munmap_calls += 1;
1105	stats.munmap_storage_request += rsize;
1106	stats.munmap_storage_alloc += size;
1107	#endif // __STATISTICS__
1108
1109	// OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
1110
1111	// Does not matter where this storage is freed.
1112	if ( unlikely( munmap( header, size ) == -1 ) ) {
1113	// Do not call strerror( errno ) as it may call malloc.
1114	abort( "** Error ** attempt to deallocate large object %p and munmap failed with errno %d.\n"
1115	"Possible cause is invalid delete pointer: either not allocated or with corrupt header.",
1116	addr, errno );
1117	} // if
1118	} else {
1119	#ifdef __CFA_DEBUG__
1120	// memset is NOT always inlined!
1121	disable_interrupts();
1122	// Scrub old memory so subsequent usages might fail. Only scrub the first/last SCRUB_SIZE bytes.
1123	char * data = ((Heap.Storage *)header)->data; // data address
1124	size_t dsize = size - sizeof(Heap.Storage); // data size
1125	if ( dsize <= SCRUB_SIZE * 2 ) {
1126	memset( data, SCRUB, dsize ); // scrub all
1127	} else {
1128	memset( data, SCRUB, SCRUB_SIZE ); // scrub front
1129	memset( data + dsize - SCRUB_SIZE, SCRUB, SCRUB_SIZE ); // scrub back
1130	} // if
1131	enable_interrupts( false );
1132	#endif // __CFA_DEBUG__
1133
1134	#ifdef OWNERSHIP
1135	if ( likely( heapManager == freeHead->homeManager ) ) { // belongs to this thread
1136	header->kind.real.next = freeHead->freeList; // push on stack
1137	freeHead->freeList = (Heap.Storage *)header;
1138	} else { // return to thread owner
1139	verify( heapManager );
1140
1141	#ifdef RETURNSPIN
1142	lock( freeHead->returnLock );
1143	header->kind.real.next = freeHead->returnList; // push to bucket return list
1144	freeHead->returnList = (Heap.Storage *)header;
1145	unlock( freeHead->returnLock );
1146	#else // lock free
1147	header->kind.real.next = freeHead->returnList; // link new node to top node
1148	// CAS resets header->kind.real.next = freeHead->returnList on failure
1149	while ( ! __atomic_compare_exchange_n( &freeHead->returnList, &header->kind.real.next, (Heap.Storage *)header,
1150	false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST ) );
1151
1152	#ifdef __STATISTICS__
1153	stats.return_pushes += 1;
1154	stats.return_storage_request += rsize;
1155	stats.return_storage_alloc += size;
1156	#endif // __STATISTICS__
1157	#endif // RETURNSPIN
1158	} // if
1159
1160	#else // no OWNERSHIP
1161
1162	// kind.real.home is address in owner thread's freeLists, so compute the equivalent position in this thread's freeList.
1163	freeHead = &freeLists[ClearStickyBits( (Heap.FreeHeader *)(header->kind.real.home) ) - &freeHead->homeManager->freeLists[0]];
1164	header->kind.real.next = freeHead->freeList; // push on stack
1165	freeHead->freeList = (Heap.Storage *)header;
1166	#endif // ! OWNERSHIP
1167
1168	// OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
1169	} // if
1170
1171	#ifdef __CFA_DEBUG__
1172	if ( traceHeap() ) {
1173	char helpText[64];
1174	__cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
1175	"Free( %p ) size:%zu\n", addr, size ); // print debug/nodebug
1176	} // if
1177	#endif // __CFA_DEBUG__
1178
1179	// poll_interrupts(); // call rollforward
1180	} // doFree
1181
1182
1183	size_t prtFree( Heap & manager ) with( manager ) {
1184	size_t total = 0;
1185	#ifdef __STATISTICS__
1186	__cfaabi_bits_acquire();
1187	__cfaabi_bits_print_nolock( STDERR_FILENO, "\nBin lists (bin size : free blocks on list)\n" );
1188	#endif // __STATISTICS__
1189	for ( unsigned int i = 0; i < heapMaster.maxBucketsUsed; i += 1 ) {
1190	size_t size = freeLists[i].blockSize;
1191	#ifdef __STATISTICS__
1192	unsigned int N = 0;
1193	#endif // __STATISTICS__
1194
1195	for ( Heap.Storage * p = freeLists[i].freeList; p != 0p; p = p->header.kind.real.next ) {
1196	total += size;
1197	#ifdef __STATISTICS__
1198	N += 1;
1199	#endif // __STATISTICS__
1200	} // for
1201
1202	#ifdef __STATISTICS__
1203	__cfaabi_bits_print_nolock( STDERR_FILENO, "%7zu, %-7u ", size, N );
1204	if ( (i + 1) % 8 == 0 ) __cfaabi_bits_print_nolock( STDERR_FILENO, "\n" );
1205	#endif // __STATISTICS__
1206	} // for
1207	#ifdef __STATISTICS__
1208	__cfaabi_bits_print_nolock( STDERR_FILENO, "\ntotal free blocks:%zu\n", total );
1209	__cfaabi_bits_release();
1210	#endif // __STATISTICS__
1211	return (char )heapMaster.heapEnd - (char )heapMaster.heapBegin - total;
1212	} // prtFree
1213
1214
1215	#ifdef __STATISTICS__
1216	static void incCalls( size_t statName ) libcfa_nopreempt {
1217	heapManager->stats.counters[statName].calls += 1;
1218	} // incCalls
1219
1220	static void incZeroCalls( size_t statName ) libcfa_nopreempt {
1221	heapManager->stats.counters[statName].calls_0 += 1;
1222	} // incZeroCalls
1223	#endif // __STATISTICS__
1224
1225	#ifdef __CFA_DEBUG__
1226	static void incUnfreed( intptr_t offset ) libcfa_nopreempt {
1227	heapManager->allocUnfreed += offset;
1228	} // incUnfreed
1229	#endif // __CFA_DEBUG__
1230
1231
1232	static void * memalignNoStats( size_t alignment, size_t size STAT_PARM ) {
1233	checkAlign( alignment ); // check alignment
1234
1235	// if alignment <= default alignment or size == 0, do normal malloc as two headers are unnecessary
1236	if ( unlikely( alignment <= libAlign() \|\| size == 0 ) ) return doMalloc( size STAT_ARG( STAT_NAME ) );
1237
1238	// Allocate enough storage to guarantee an address on the alignment boundary, and sufficient space before it for
1239	// administrative storage. NOTE, WHILE THERE ARE 2 HEADERS, THE FIRST ONE IS IMPLICITLY CREATED BY DOMALLOC.
1240	// .-------------v-----------------v----------------v----------,
1241	// \| Real Header \| ... padding ... \| Fake Header \| data ... \|
1242	// `-------------^-----------------^-+--------------^----------'
1243	// \|<--------------------------------' offset/align \|<-- alignment boundary
1244
1245	// subtract libAlign() because it is already the minimum alignment
1246	// add sizeof(Storage) for fake header
1247	size_t offset = alignment - libAlign() + sizeof(Heap.Storage);
1248	char * addr = (char *)doMalloc( size + offset STAT_ARG( STAT_NAME ) );
1249
1250	// address in the block of the "next" alignment address
1251	char * user = (char *)ceiling2( (uintptr_t)(addr + sizeof(Heap.Storage)), alignment );
1252
1253	// address of header from malloc
1254	Heap.Storage.Header * realHeader = HeaderAddr( addr );
1255	realHeader->kind.real.size = size; // correct size to eliminate above alignment offset
1256	#ifdef __CFA_DEBUG__
1257	incUnfreed( -offset ); // adjustment off the offset from call to doMalloc
1258	#endif // __CFA_DEBUG__
1259
1260	// address of fake header before the alignment location
1261	Heap.Storage.Header * fakeHeader = HeaderAddr( user );
1262
1263	// SKULLDUGGERY: insert the offset to the start of the actual storage block and remember alignment
1264	fakeHeader->kind.fake.offset = (char )fakeHeader - (char )realHeader;
1265	// SKULLDUGGERY: odd alignment implies fake header
1266	fakeHeader->kind.fake.alignment = MarkAlignmentBit( alignment );
1267
1268	return user;
1269	} // memalignNoStats
1270
1271
1272	//####################### Memory Allocation Routines ####################
1273
1274
1275	extern "C" {
1276	// Allocates size bytes and returns a pointer to the allocated memory. The contents are undefined. If size is 0,
1277	// then malloc() returns a unique pointer value that can later be successfully passed to free().
1278	void * malloc( size_t size ) libcfa_public {
1279	return doMalloc( size STAT_ARG( MALLOC ) );
1280	} // malloc
1281
1282
1283	// Same as malloc() except size bytes is an array of dim elements each of elemSize bytes.
1284	void * aalloc( size_t dim, size_t elemSize ) libcfa_public {
1285	return doMalloc( dim * elemSize STAT_ARG( AALLOC ) );
1286	} // aalloc
1287
1288
1289	// Same as aalloc() with memory set to zero.
1290	void * calloc( size_t dim, size_t elemSize ) libcfa_public {
1291	size_t size = dim * elemSize;
1292	char * addr = (char *)doMalloc( size STAT_ARG( CALLOC ) );
1293
1294	if ( unlikely( addr == NULL ) ) return NULL; // stop further processing if 0p is returned
1295
1296	Heap.Storage.Header * header;
1297	Heap.FreeHeader * freeHead;
1298	size_t bsize, alignment;
1299
1300	#ifndef __CFA_DEBUG__
1301	bool mapped =
1302	#endif // __CFA_DEBUG__
1303	headers( "calloc", addr, header, freeHead, bsize, alignment );
1304
1305	#ifndef __CFA_DEBUG__
1306	// Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
1307	if ( likely( ! mapped ) )
1308	#endif // __CFA_DEBUG__
1309	// <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
1310	// `-header`-addr `-size
1311	memset( addr, '\0', size ); // set to zeros
1312
1313	MarkZeroFilledBit( header ); // mark as zero fill
1314	return addr;
1315	} // calloc
1316
1317
1318	// Change the size of the memory block pointed to by oaddr to size bytes. The contents are undefined. If oaddr is
1319	// 0p, then the call is equivalent to malloc(size), for all values of size; if size is equal to zero, and oaddr is
1320	// not 0p, then the call is equivalent to free(oaddr). Unless oaddr is 0p, it must have been returned by an earlier
1321	// call to malloc(), alloc(), calloc() or realloc(). If the area pointed to was moved, a free(oaddr) is done.
1322	void * resize( void * oaddr, size_t size ) libcfa_public {
1323	if ( unlikely( oaddr == 0p ) ) { // => malloc( size )
1324	return doMalloc( size STAT_ARG( RESIZE ) );
1325	} // if
1326
1327	PROLOG( RESIZE, doFree( oaddr ) ); // => free( oaddr )
1328
1329	Heap.Storage.Header * header;
1330	Heap.FreeHeader * freeHead;
1331	size_t bsize, oalign;
1332	headers( "resize", oaddr, header, freeHead, bsize, oalign );
1333
1334	size_t odsize = DataStorage( bsize, oaddr, header ); // data storage available in bucket
1335	// same size, DO NOT preserve STICKY PROPERTIES.
1336	if ( oalign == libAlign() && size <= odsize && odsize <= size * 2 ) { // allow 50% wasted storage for smaller size
1337	ClearZeroFillBit( header ); // no alignment and turn off 0 fill
1338	#ifdef __CFA_DEBUG__
1339	incUnfreed( size - header->kind.real.size ); // adjustment off the size difference
1340	#endif // __CFA_DEBUG__
1341	header->kind.real.size = size; // reset allocation size
1342	#ifdef __STATISTICS__
1343	incCalls( RESIZE );
1344	#endif // __STATISTICS__
1345	return oaddr;
1346	} // if
1347
1348	// change size, DO NOT preserve STICKY PROPERTIES.
1349	doFree( oaddr ); // free previous storage
1350
1351	return doMalloc( size STAT_ARG( RESIZE ) ); // create new area
1352	} // resize
1353
1354
1355	// Same as resize() but the contents are unchanged in the range from the start of the region up to the minimum of
1356	// the old and new sizes.
1357	void * realloc( void * oaddr, size_t size ) libcfa_public {
1358	if ( unlikely( oaddr == 0p ) ) { // => malloc( size )
1359	return doMalloc( size STAT_ARG( REALLOC ) );
1360	} // if
1361
1362	PROLOG( REALLOC, doFree( oaddr ) ); // => free( oaddr )
1363
1364	Heap.Storage.Header * header;
1365	Heap.FreeHeader * freeHead;
1366	size_t bsize, oalign;
1367	headers( "realloc", oaddr, header, freeHead, bsize, oalign );
1368
1369	size_t odsize = DataStorage( bsize, oaddr, header ); // data storage available in bucket
1370	size_t osize = header->kind.real.size; // old allocation size
1371	bool ozfill = ZeroFillBit( header ); // old allocation zero filled
1372	if ( unlikely( size <= odsize ) && odsize <= size * 2 ) { // allow up to 50% wasted storage
1373	#ifdef __CFA_DEBUG__
1374	incUnfreed( size - header->kind.real.size ); // adjustment off the size difference
1375	#endif // __CFA_DEBUG__
1376	header->kind.real.size = size; // reset allocation size
1377	if ( unlikely( ozfill ) && size > osize ) { // previous request zero fill and larger ?
1378	memset( (char *)oaddr + osize, '\0', size - osize ); // initialize added storage
1379	} // if
1380	#ifdef __STATISTICS__
1381	incCalls( REALLOC );
1382	#endif // __STATISTICS__
1383	return oaddr;
1384	} // if
1385
1386	// change size and copy old content to new storage
1387
1388	void * naddr;
1389	if ( likely( oalign <= libAlign() ) ) { // previous request not aligned ?
1390	naddr = doMalloc( size STAT_ARG( REALLOC ) ); // create new area
1391	} else {
1392	naddr = memalignNoStats( oalign, size STAT_ARG( REALLOC ) ); // create new aligned area
1393	} // if
1394
1395	headers( "realloc", naddr, header, freeHead, bsize, oalign );
1396	// To preserve prior fill, the entire bucket must be copied versus the size.
1397	memcpy( naddr, oaddr, min( osize, size ) ); // copy bytes
1398	doFree( oaddr ); // free previous storage
1399
1400	if ( unlikely( ozfill ) ) { // previous request zero fill ?
1401	MarkZeroFilledBit( header ); // mark new request as zero filled
1402	if ( size > osize ) { // previous request larger ?
1403	memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
1404	} // if
1405	} // if
1406	return naddr;
1407	} // realloc
1408
1409
1410	// Same as realloc() except the new allocation size is large enough for an array of nelem elements of size elsize.
1411	void * reallocarray( void * oaddr, size_t dim, size_t elemSize ) libcfa_public {
1412	return realloc( oaddr, dim * elemSize );
1413	} // reallocarray
1414
1415
1416	// Same as malloc() except the memory address is a multiple of alignment, which must be a power of two. (obsolete)
1417	void * memalign( size_t alignment, size_t size ) libcfa_public {
1418	return memalignNoStats( alignment, size STAT_ARG( MEMALIGN ) );
1419	} // memalign
1420
1421
1422	// Same as aalloc() with memory alignment.
1423	void * amemalign( size_t alignment, size_t dim, size_t elemSize ) libcfa_public {
1424	return memalignNoStats( alignment, dim * elemSize STAT_ARG( AMEMALIGN ) );
1425	} // amemalign
1426
1427
1428	// Same as calloc() with memory alignment.
1429	void * cmemalign( size_t alignment, size_t dim, size_t elemSize ) libcfa_public {
1430	size_t size = dim * elemSize;
1431	char * addr = (char *)memalignNoStats( alignment, size STAT_ARG( CMEMALIGN ) );
1432
1433	if ( unlikely( addr == NULL ) ) return NULL; // stop further processing if 0p is returned
1434
1435	Heap.Storage.Header * header;
1436	Heap.FreeHeader * freeHead;
1437	size_t bsize;
1438
1439	#ifndef __CFA_DEBUG__
1440	bool mapped =
1441	#endif // __CFA_DEBUG__
1442	headers( "cmemalign", addr, header, freeHead, bsize, alignment );
1443
1444	// Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
1445	#ifndef __CFA_DEBUG__
1446	if ( ! mapped )
1447	#endif // __CFA_DEBUG__
1448	// <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
1449	// `-header`-addr `-size
1450	memset( addr, '\0', size ); // set to zeros
1451
1452	MarkZeroFilledBit( header ); // mark as zero filled
1453	return addr;
1454	} // cmemalign
1455
1456
1457	// Same as memalign(), but ISO/IEC 2011 C11 Section 7.22.2 states: the value of size shall be an integral multiple
1458	// of alignment. This requirement is universally ignored.
1459	void * aligned_alloc( size_t alignment, size_t size ) libcfa_public {
1460	return memalign( alignment, size );
1461	} // aligned_alloc
1462
1463
1464	// Allocates size bytes and places the address of the allocated memory in *memptr. The address of the allocated
1465	// memory shall be a multiple of alignment, which must be a power of two and a multiple of sizeof(void *). If size
1466	// is 0, then posix_memalign() returns either 0p, or a unique pointer value that can later be successfully passed to
1467	// free(3).
1468	int posix_memalign( void ** memptr, size_t alignment, size_t size ) libcfa_public {
1469	if ( unlikely( alignment < libAlign() \|\| ! is_pow2( alignment ) ) ) return EINVAL; // check alignment
1470	*memptr = memalign( alignment, size );
1471	return 0;
1472	} // posix_memalign
1473
1474
1475	// Allocates size bytes and returns a pointer to the allocated memory. The memory address shall be a multiple of the
1476	// page size. It is equivalent to memalign(sysconf(_SC_PAGESIZE),size).
1477	void * valloc( size_t size ) libcfa_public {
1478	return memalign( __page_size, size );
1479	} // valloc
1480
1481
1482	// Same as valloc but rounds size to multiple of page size.
1483	void * pvalloc( size_t size ) libcfa_public {
1484	return memalign( __page_size, ceiling2( size, __page_size ) ); // round size to multiple of page size
1485	} // pvalloc
1486
1487
1488	// Frees the memory space pointed to by ptr, which must have been returned by a previous call to malloc(), calloc()
1489	// or realloc(). Otherwise, or if free(ptr) has already been called before, undefined behaviour occurs. If ptr is
1490	// 0p, no operation is performed.
1491	void free( void * addr ) libcfa_public {
1492	if ( unlikely( addr == 0p ) ) { // special case
1493	#ifdef __STATISTICS__
1494	if ( heapManager )
1495	incZeroCalls( FREE );
1496	#endif // __STATISTICS__
1497	return;
1498	} // if
1499
1500	#ifdef __STATISTICS__
1501	incCalls( FREE );
1502	#endif // __STATISTICS__
1503
1504	doFree( addr ); // handles heapManager == nullptr
1505	} // free
1506
1507
1508	// Returns the alignment of an allocation.
1509	size_t malloc_alignment( void * addr ) libcfa_public {
1510	if ( unlikely( addr == 0p ) ) return libAlign(); // minimum alignment
1511	Heap.Storage.Header * header = HeaderAddr( addr );
1512	if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
1513	return ClearAlignmentBit( header ); // clear flag from value
1514	} else {
1515	return libAlign(); // minimum alignment
1516	} // if
1517	} // malloc_alignment
1518
1519
1520	// Returns true if the allocation is zero filled, e.g., allocated by calloc().
1521	bool malloc_zero_fill( void * addr ) libcfa_public {
1522	if ( unlikely( addr == 0p ) ) return false; // null allocation is not zero fill
1523	Heap.Storage.Header * header = HeaderAddr( addr );
1524	if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
1525	header = RealHeader( header ); // backup from fake to real header
1526	} // if
1527	return ZeroFillBit( header ); // zero filled ?
1528	} // malloc_zero_fill
1529
1530
1531	// Returns original total allocation size (not bucket size) => array size is dimension * sizeof(T).
1532	size_t malloc_size( void * addr ) libcfa_public {
1533	if ( unlikely( addr == 0p ) ) return 0; // null allocation has zero size
1534	Heap.Storage.Header * header = HeaderAddr( addr );
1535	if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
1536	header = RealHeader( header ); // backup from fake to real header
1537	} // if
1538	return header->kind.real.size;
1539	} // malloc_size
1540
1541
1542	// Returns the number of usable bytes in the block pointed to by ptr, a pointer to a block of memory allocated by
1543	// malloc or a related function.
1544	size_t malloc_usable_size( void * addr ) libcfa_public {
1545	if ( unlikely( addr == 0p ) ) return 0; // null allocation has 0 size
1546	Heap.Storage.Header * header;
1547	Heap.FreeHeader * freeHead;
1548	size_t bsize, alignment;
1549
1550	headers( "malloc_usable_size", addr, header, freeHead, bsize, alignment );
1551	return DataStorage( bsize, addr, header ); // data storage in bucket
1552	} // malloc_usable_size
1553
1554
1555	// Prints (on default standard error) statistics about memory allocated by malloc and related functions.
1556	void malloc_stats( void ) libcfa_public {
1557	#ifdef __STATISTICS__
1558	HeapStatistics stats;
1559	HeapStatisticsCtor( stats );
1560	if ( printStats( collectStats( stats ) ) == -1 ) {
1561	#else
1562	#define MALLOC_STATS_MSG "malloc_stats statistics disabled.\n"
1563	if ( write( STDERR_FILENO, MALLOC_STATS_MSG, sizeof( MALLOC_STATS_MSG ) - 1 /* size includes '\0' */ ) == -1 ) {
1564	#endif // __STATISTICS__
1565	abort( "** Error ** write failed in malloc_stats" );
1566	} // if
1567	} // malloc_stats
1568
1569
1570	// Zero the heap master and all active thread heaps.
1571	void malloc_stats_clear() {
1572	#ifdef __STATISTICS__
1573	clearStats();
1574	#else
1575	#define MALLOC_STATS_MSG "malloc_stats statistics disabled.\n"
1576	if ( write( STDERR_FILENO, MALLOC_STATS_MSG, sizeof( MALLOC_STATS_MSG ) - 1 /* size includes '\0' */ ) == -1 ) {
1577	abort( "** Error ** write failed in malloc_stats" );
1578	} // if
1579	#endif // __STATISTICS__
1580	} // malloc_stats_clear
1581
1582
1583	// Changes the file descriptor where malloc_stats() writes statistics.
1584	int malloc_stats_fd( int fd __attribute__(( unused )) ) libcfa_public {
1585	#ifdef __STATISTICS__
1586	int temp = heapMaster.stats_fd;
1587	heapMaster.stats_fd = fd;
1588	return temp;
1589	#else
1590	return -1; // unsupported
1591	#endif // __STATISTICS__
1592	} // malloc_stats_fd
1593
1594
1595	// Prints an XML string that describes the current state of the memory-allocation implementation in the caller.
1596	// The string is printed on the file stream stream. The exported string includes information about all arenas (see
1597	// malloc).
1598	int malloc_info( int options, FILE * stream __attribute__(( unused )) ) libcfa_public {
1599	if ( options != 0 ) { errno = EINVAL; return -1; }
1600	#ifdef __STATISTICS__
1601	HeapStatistics stats;
1602	HeapStatisticsCtor( stats );
1603	return printStatsXML( collectStats( stats ), stream ); // returns bytes written or -1
1604	#else
1605	return 0; // unsupported
1606	#endif // __STATISTICS__
1607	} // malloc_info
1608
1609
1610	// Adjusts parameters that control the behaviour of the memory-allocation functions (see malloc). The param argument
1611	// specifies the parameter to be modified, and value specifies the new value for that parameter.
1612	int mallopt( int option, int value ) libcfa_public {
1613	if ( value < 0 ) return 0;
1614	choose( option ) {
1615	case M_TOP_PAD:
1616	heapMaster.heapExpand = ceiling2( value, __page_size );
1617	return 1;
1618	case M_MMAP_THRESHOLD:
1619	if ( setMmapStart( value ) ) return 1;
1620	} // choose
1621	return 0; // error, unsupported
1622	} // mallopt
1623
1624
1625	// Attempt to release free memory at the top of the heap (by calling sbrk with a suitable argument).
1626	int malloc_trim( size_t ) libcfa_public {
1627	return 0; // => impossible to release memory
1628	} // malloc_trim
1629
1630
1631	// Records the current state of all malloc internal bookkeeping variables (but not the actual contents of the heap
1632	// or the state of malloc_hook functions pointers). The state is recorded in a system-dependent opaque data
1633	// structure dynamically allocated via malloc, and a pointer to that data structure is returned as the function
1634	// result. (The caller must free this memory.)
1635	void * malloc_get_state( void ) libcfa_public {
1636	return 0p; // unsupported
1637	} // malloc_get_state
1638
1639
1640	// Restores the state of all malloc internal bookkeeping variables to the values recorded in the opaque data
1641	// structure pointed to by state.
1642	int malloc_set_state( void * ) libcfa_public {
1643	return 0; // unsupported
1644	} // malloc_set_state
1645
1646
1647	// Sets the amount (bytes) to extend the heap when there is insufficent free storage to service an allocation.
1648	__attribute__((weak)) size_t malloc_expansion() libcfa_public { return __CFA_DEFAULT_HEAP_EXPANSION__; }
1649
1650	// Sets the crossover point between allocations occuring in the sbrk area or separately mmapped.
1651	__attribute__((weak)) size_t malloc_mmap_start() libcfa_public { return __CFA_DEFAULT_MMAP_START__; }
1652
1653	// Amount subtracted to adjust for unfreed program storage (debug only).
1654	__attribute__((weak)) size_t malloc_unfreed() libcfa_public { return __CFA_DEFAULT_HEAP_UNFREED__; }
1655	} // extern "C"
1656
1657
1658	// Must have CFA linkage to overload with C linkage realloc.
1659	void * resize( void * oaddr, size_t nalign, size_t size ) libcfa_public {
1660	if ( unlikely( oaddr == 0p ) ) { // => malloc( size )
1661	return memalignNoStats( nalign, size STAT_ARG( RESIZE ) );
1662	} // if
1663
1664	PROLOG( RESIZE, doFree( oaddr ) ); // => free( oaddr )
1665
1666	// Attempt to reuse existing alignment.
1667	Heap.Storage.Header * header = HeaderAddr( oaddr );
1668	bool isFakeHeader = AlignmentBit( header ); // old fake header ?
1669	size_t oalign;
1670
1671	if ( unlikely( isFakeHeader ) ) {
1672	checkAlign( nalign ); // check alignment
1673	oalign = ClearAlignmentBit( header ); // old alignment
1674	if ( unlikely( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1675	&& ( oalign <= nalign // going down
1676	\|\| (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1677	) ) {
1678	HeaderAddr( oaddr )->kind.fake.alignment = MarkAlignmentBit( nalign ); // update alignment (could be the same)
1679	Heap.FreeHeader * freeHead;
1680	size_t bsize, oalign;
1681	headers( "resize", oaddr, header, freeHead, bsize, oalign );
1682	size_t odsize = DataStorage( bsize, oaddr, header ); // data storage available in bucket
1683
1684	if ( size <= odsize && odsize <= size * 2 ) { // allow 50% wasted data storage
1685	HeaderAddr( oaddr )->kind.fake.alignment = MarkAlignmentBit( nalign ); // update alignment (could be the same)
1686	ClearZeroFillBit( header ); // turn off 0 fill
1687	#ifdef __CFA_DEBUG__
1688	incUnfreed( size - header->kind.real.size ); // adjustment off the size difference
1689	#endif // __CFA_DEBUG__
1690	header->kind.real.size = size; // reset allocation size
1691	#ifdef __STATISTICS__
1692	incCalls( RESIZE );
1693	#endif // __STATISTICS__
1694	return oaddr;
1695	} // if
1696	} // if
1697	} else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1698	&& nalign == libAlign() ) { // new alignment also on libAlign => no fake header needed
1699	return resize( oaddr, size ); // duplicate special case checks
1700	} // if
1701
1702	// change size, DO NOT preserve STICKY PROPERTIES.
1703	doFree( oaddr ); // free previous storage
1704	return memalignNoStats( nalign, size STAT_ARG( RESIZE ) ); // create new aligned area
1705	} // resize
1706
1707
1708	void * realloc( void * oaddr, size_t nalign, size_t size ) libcfa_public {
1709	if ( unlikely( oaddr == 0p ) ) { // => malloc( size )
1710	return memalignNoStats( nalign, size STAT_ARG( REALLOC ) );
1711	} // if
1712
1713	PROLOG( REALLOC, doFree( oaddr ) ); // => free( oaddr )
1714
1715	// Attempt to reuse existing alignment.
1716	Heap.Storage.Header * header = HeaderAddr( oaddr );
1717	bool isFakeHeader = AlignmentBit( header ); // old fake header ?
1718	size_t oalign;
1719	if ( unlikely( isFakeHeader ) ) {
1720	checkAlign( nalign ); // check alignment
1721	oalign = ClearAlignmentBit( header ); // old alignment
1722	if ( unlikely( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1723	&& ( oalign <= nalign // going down
1724	\|\| (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1725	) ) {
1726	HeaderAddr( oaddr )->kind.fake.alignment = MarkAlignmentBit( nalign ); // update alignment (could be the same)
1727	return realloc( oaddr, size ); // duplicate special case checks
1728	} // if
1729	} else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1730	&& nalign == libAlign() ) { // new alignment also on libAlign => no fake header needed
1731	return realloc( oaddr, size ); // duplicate special case checks
1732	} // if
1733
1734	Heap.FreeHeader * freeHead;
1735	size_t bsize;
1736	headers( "realloc", oaddr, header, freeHead, bsize, oalign );
1737
1738	// change size and copy old content to new storage
1739
1740	size_t osize = header->kind.real.size; // old allocation size
1741	bool ozfill = ZeroFillBit( header ); // old allocation zero filled
1742
1743	void * naddr = memalignNoStats( nalign, size STAT_ARG( REALLOC ) ); // create new aligned area
1744
1745	headers( "realloc", naddr, header, freeHead, bsize, oalign );
1746	memcpy( naddr, oaddr, min( osize, size ) ); // copy bytes
1747	doFree( oaddr ); // free previous storage
1748
1749	if ( unlikely( ozfill ) ) { // previous request zero fill ?
1750	MarkZeroFilledBit( header ); // mark new request as zero filled
1751	if ( size > osize ) { // previous request larger ?
1752	memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
1753	} // if
1754	} // if
1755	return naddr;
1756	} // realloc
1757
1758
1759	void * reallocarray( void * oaddr, size_t nalign, size_t dim, size_t elemSize ) __THROW {
1760	return realloc( oaddr, nalign, dim * elemSize );
1761	} // reallocarray
1762
1763
1764	// Local Variables: //
1765	// tab-width: 4 //
1766	// compile-command: "cfa -nodebug -O2 heap.cfa" //
1767	// End: //

Note: See TracBrowser for help on using the repository browser.

Download in other formats:

Original Format